Source-linked AI summary
Detoxifying Toxic Communication: A Design Science Approach to Responsible AI
Hossein Arshadi Soufiani, Henry M. Kim, Hjalmar Turesson, Syed Mohammad Arham Noman, Anav Setia
TL;DR
Toxic workplace communication can undermine trust and collaboration, while existing moderation often deletes or blocks messages rather than resolving them constructively. This study develops a Design Science Research artifact combining toxicity detection with generative detoxification for organizational communication. The artifact achieves strong detection performance and semantic preservation, supporting respectful communication while retaining meaning and continuity.
Problem
Existing moderation tools emphasize detection, deletion, or flagging, while workplace detoxification lacks guidance addressing semantic preservation, tone stability, conversation continuity, fairness, and user acceptance.
Method
The study uses Design Science Research to develop a responsible AI artifact integrating fine-tuned transformer classifiers with controlled generative detoxification for organizational communication.
Results
Accuracy of 0.965 and F1-score of 0.979 were achieved for toxicity classification, while detoxified outputs reached a mean F_BERT score of 0.930.
Takeaways & Limitations
The artifact demonstrates that toxic discourse can be constructively transformed while preserving semantic meaning and conversational continuity in digital workplaces.
Takeaways & Limitations
Fairness auditing was limited by absent demographic data, while multilingual and informal performance, computational constraints, and dataset limitations restricted generalizability.
Abstract
from arXiv · showhide
Toxic language in digital workplaces such as pejoratives, sarcasm, condescension, and subtle incivility can erode trust, morale, and collaboration. Existing moderation tools primarily delete or block harmful messages, disrupting communication and offering no constructive resolution. This study adopts a Design Science Research approach to create a responsible AI artifact that detects and detoxifies toxic communication. The artifact integrates fine-tuned transformer-based classifiers (DistilBERT, DistilRoBERTa) with a generative detoxification model (mT0-XL-Detox-ORPO) that rewrites toxic text into semantically equivalent, non-offensive paraphrases. Technical evaluation demonstrates high accuracy in toxicity detection and strong semantic preservation in rewritten messages, supporting conversation continuity while reinforcing respectful discourse. The paper contributes design principles for responsible AI moderation that prioritize meaning preservation and fairness.
1. Introduction
Digital workplace platforms support collaboration but also enable subtle toxic communication that can erode organizational culture. The study proposes a Design Science Research artifact that detects and transforms toxic messages while preserving communication continuity.
- 1. Introduction: Workplace communication platforms enable rapid collaboration across distributed organizations but also provide channels for subtle incivility.Examples include sarcasm, condescending comments, and negative gossip that may avoid overt slurs.
- 1. Introduction: Online disinhibition and limited face-to-face cues can intensify toxic behavior in asynchronous text communication.Minor norm violations may accumulate, erode organizational culture, and spiral into more serious conflicts if unchecked.
- 1. Introduction: Detoxification offers a way to improve digital discourse without erasing the speaker’s voice.The approach reframes moderation as constructive revision rather than removal.
- 1. Introduction: Prior work had not developed or evaluated a Design Science Research artifact that automatically detoxifies workplace communication while explicitly operationalizing responsible AI principles.Existing tools emphasize detection, deletion, or flagging and rarely address semantic preservation, tone stability, and conversation continuity in organizational settings.
- 1. Introduction: The paper adopts Design Science Research to develop a responsible AI artifact for moderating toxic language in organizational communication channels.The paper presents the artifact’s design, technical evaluation, implications, ethical considerations, and limitations.
2. Literature Review
Prior research shows that toxicity detection and detoxification remain difficult because meaning, context, fairness, transparency, and organizational requirements must be handled together. This study uses Design Science Research to translate those requirements into a responsible, implementable artifact.
- 2. Literature Review: Toxicity classifiers can miss context-dependent incivility because acceptability depends on subtle cues, conversation history, and organizational culture.Off-the-shelf systems often rely on narrow definitions focused on insults or hate speech.
- 2. Literature Review: Automated moderation can reproduce cultural bias, unfairly target marginalized communities, and obscure why messages are flagged or altered.Enterprise tools therefore need domain sensitivity, professional tone handling, and fair treatment.
- 2. Literature Review: Toxicity datasets often contain fewer than 10% toxic examples, which can bias classifiers toward predicting neutral content and reduce recall.This imbalance limits coverage of genuinely toxic incidents.
- 2. Literature Review: Sequence-to-sequence transformers can rewrite toxic text into semantically equivalent, non-toxic phrasing, supporting revision over censorship.Beam search, semantic scoring, and re-ranking have been used to improve generative detoxification.
- 2. Literature Review: Highly accurate moderation models may still make false toxicity accusations, while large-model costs and latency complicate deployment.Appeal mechanisms and justification are relevant to perceived legitimacy and bias.
- 2. Literature Review: Detoxification is a specialized form of style transfer requiring especially strong meaning preservation for organizational communication.Factual details, intent, and requests should remain intact after transformation.
- 2. Literature Review: Design Science Research translates organizational requirements into concrete design requirements and an implementable responsible AI artifact.This addresses the limited guidance on designing detoxification for workplace communication, where user acceptance is critical.
3. Artifact Design Development
The artifact applies Design Science Research to build a modular, responsible AI pipeline that detects toxic organizational communication and rewrites flagged content while preserving meaning. Iterative evaluation combines classification performance with semantic-preservation and detoxification measures, while planned organizational deployment and feedback remain future work.
- Design Science Research: The DSR methodology structures artifact development through iterative problem identification, objective definition, design, demonstration, evaluation, and communication activities.The process was implemented reflexively, with iteration insights guiding model configuration and dataset adaptation.
- Design Requirements: The design prioritizes accurate contextual toxicity detection, semantically faithful detoxification, and modularity for further experimentation.Content preservation and tone consistency are instantiated through side-by-side text, semantic similarity scoring, and candidate re-ranking.
- System Architecture: The detect-then-mitigate architecture separates toxicity classification from generative rewriting of flagged content.This modular structure enables separate tuning of intervention thresholds and detoxification behavior.
- Scope and Future Work: Real-time enterprise integration and user-feedback components are planned extensions rather than implemented parts of the present artifact.Future evaluation is intended to examine trust, fairness perception, and behavioral adaptation after deployment through enterprise API connectors.
- Technical Performance: DistilBERT and DistilRoBERTa classifiers achieved accuracy of 0.965 and F1-score of 0.979 on the adapted Davidson et al. dataset.The models were selected with efficiency in mind for small and medium-sized organizations with limited infrastructure.
- Technical Performance: Detoxified outputs achieved a mean F_BERT score of 0.930, with P_BERT = 0.969 and R_BERT = 0.948, indicating strong semantic preservation.Style Transfer Accuracy was also used to verify that rewritten texts were recognized as non-toxic by an independent classifier.
4. Discussion
The artifact embeds responsible AI principles into organizational communication moderation, combining detection and detoxification while preserving clarity and workflow continuity. The discussion also identifies fairness, multilingual performance, dataset diversity, transparency, and computational constraints as unresolved challenges.
- The artifact integrates toxic-language detection and detoxification as a socio-technical solution for organizational communication.It is intended to enhance communication ethics without losing clarity or disrupting workflows.
- Iterative DSR refinement addressed early weaknesses in semantic preservation and fairness through model tuning and user feedback.
- Detoxification performance declined in multilingual or informal environments, while fairness auditing was constrained by missing enterprise demographic data.
- Pre-existing datasets and benchmarks lacked sufficient dialectal and identity-based diversity, limiting fairness evaluation and cross-cultural generalizability.
- Deployment requires balancing fluency, meaning preservation, transparency, computational complexity, organizational readiness, communication policies, and stakeholder training.
5. Concluding Remarks
The study presents a Design Science Research approach for a responsible AI artifact that detects and detoxifies toxic organizational communication while preserving semantic meaning. It identifies limitations in fairness auditing, generalizability, multilingual and informal performance, and scalability, and proposes human-in-the-loop, governance, and domain-specific extensions.
- The study uses Design Science Research to develop a responsible AI artifact for detecting and detoxifying toxic organizational communication.The artifact integrates large language models into a modular framework.
- The artifact demonstrates that toxic discourse can be constructively transformed while preserving semantic meaning.
- Fairness auditing was limited by absent demographic data, while pre-existing datasets constrained crosscultural and contextual generalizability.
- Performance in multilingual and informal environments, together with computational limitations, indicates a need for continued refinement and scalability.
- Future work should add human-in-the-loop evaluation, governance integration, and domain-specific deployment to capture real-world and longitudinal effects.
A. Appendix
The appendix presents implementation snippets for preparing datasets, configuring a DistilRoBERTa toxicity classifier, and invoking a detoxification model.
- Data preparation: The implementation loads training, validation, and test data into Hugging Face datasets.The datasets are created from pandas dataframes, with columns renamed for text and labels.
- Model components: The appendix identifies DistilRoBERTa as the classifier model and includes a separate detoxifier model for detoxification.The classifier is loaded from the distilroberta-base model identifier, while the detoxifier implementation uses language-specific prompts.
- Data preparation: The classifier uses RobertaTokenizerFast with truncation, padding, and a maximum sequence length of 256.Tokenized batches also retain the corresponding labels for model evaluation or training.
- Classifier configuration: The toxicity model is configured with two labels: not_offensive and offensive.The configuration maps numeric class identifiers to these labels and back.